D RAFT VERSION O CTOBER 3, 2017

Typeset using LATEX twocolumn style in AASTeX61

THE ELECTROMAGNETIC COUNTERPART OF THE BINARY NEUTRON STAR MERGER LIGO/VIRGO GW170817.
IV. DETECTION OF NEAR-INFRARED SIGNATURES OF r-PROCESS NUCLEOSYNTHESIS WITH GEMINI-SOUTH

R. C HORNOCK , 1 E. B ERGER , 2 D. K ASEN , 3, 4 P. S. C OWPERTHWAITE , 2 , ∗ M. N ICHOLL , 2 V. A. V ILLAR , 2 , ∗ K. D. A LEXANDER , 2

P. K. B LANCHARD , 2 T. E FTEKHARI , 2 W. F ONG , 5 , † R. M ARGUTTI , 5 P. K. G. W ILLIAMS , 2 J. A NNIS , 6 D. B ROUT , 7 D. A. B ROWN , 8
H.-Y. C HEN , 9 M. R. D ROUT , 10 , ‡ R. J. F OLEY , 11 J. A. F RIEMAN , 12, 13 C. L. F RYER , 14 D. E. H OLZ , 15 T. M ATHESON , 16
B. D. M ETZGER , 17 E. Q UATAERT , 18 A. R EST , 19, 20 M. S AKO , 7 D. M. S COLNIC , 21 N. S MITH , 22 AND M. S OARES -S ANTOS 23, 6

1 Astrophysical Institute, Department of Physics and Astronomy, 251B Clippinger Lab, Ohio University, Athens, OH 45701, USA; [email protected]
2 Harvard-Smithsonian Center for Astrophysics, 60 Garden Street, Cambridge, MA 02138, USA
3 Departments of Physics and Astronomy, and Theoretical Astrophysics Center, University of California, Berkeley, CA 94720-7300, USA
4 Nuclear Science Division, Lawrence Berkeley National Laboratory, Berkeley, CA 94720-8169, USA
5 Center for Interdisciplinary Exploration and Research in Astrophysics (CIERA) and Department of Physics and Astronomy, Northwestern University, Evanston, IL

60208
6 Fermi National Accelerator Laboratory, P. O. Box 500, Batavia, IL 60510, USA
7 Department of Physics and Astronomy, University of Pennsylvania, Philadelphia, PA 19104, USA
8 Department of Physics, Syracuse University, Syracuse NY 13224, USA
9 Department of Astronomy and Astrophysics, University of Chicago, Chicago, Illinois 60637, USA
10 The Observatories of the Carnegie Institution for Science, 813 Santa Barbara St., Pasadena, CA 91101, USA
11 Department of Astronomy and Astrophysics, University of California, Santa Cruz, CA 95064, USA

12 Fermi National Accelerator Laboratory, P. O. Box 500, Batavia, IL 60510

13 KavliInstitute for Cosmological Physics, The University of Chicago, Chicago, IL 60637
14 Center for Theoretical Astrophysics, Los Alamos National Laboratory, Los Alamos, NM 87544
15 Enrico Fermi Institute, Department of Physics, Department of Astronomy and Astrophysics,

and Kavli Institute for Cosmological Physics, University of Chicago, Chicago, IL 60637, USA
16 National Optical Astronomy Observatory, 950 North Cherry Avenue, Tucson, AZ, 85719
17 Department of Physics and Columbia Astrophysics Laboratory, Columbia University, New York, NY 10027, USA
18 Department of Astronomy & Theoretical Astrophysics Center, University of California, Berkeley, CA 94720-3411, USA
19 Space Telescope Science Institute, 3700 San Martin Drive, Baltimore, MD 21218, USA

20 Department of Physics and Astronomy, The Johns Hopkins University, 3400 North Charles Street, Baltimore, MD 21218, USA
21 Kavli Institute for Cosmological Physics, The University of Chicago, Chicago,IL 60637, USA
22 Steward Observatory, University of Arizona, 933 N. Cherry Ave., Tucson, AZ 85721
23 Department of Physics, Brandeis University, Waltham, MA 02454, USA

ABSTRACT
We present a near-infrared spectral sequence of the electromagnetic counterpart to the binary neutron star merger GW170817
detected by Advanced LIGO/Virgo. Our dataset comprises 7 epochs of J + H spectra taken with FLAMINGOS-2 on Gemini-
South between 1.5–10.5 days after the merger. In the initial epoch, the spectrum is dominated by a smooth blue continuum
due to a high-velocity, lanthanide-poor blue kilonova component. Starting the following night, all subsequent spectra instead
show features that are similar to those predicted in model spectra of material with a high concentration of lanthanides, including
spectral peaks near 1.07 and 1.55 µm. Our fiducial model with 0.04 M of ejecta, an ejection velocity of v = 0.1c, and a lanthanide
concentration of Xlan =10−2 provides a good match to the spectra taken in the first five days, although it over-predicts the late-time
fluxes. We also explore models with multiple fitting components, in each case finding that a significant abundance of lanthanide
elements is necessary to match the broad spectral peaks that we observe starting at 2.5 d after the merger. These data provide
direct evidence that binary neutron star mergers are significant production sites of even the heaviest r-process elements.

Keywords: stars: neutron — binaries: close — nuclear reactions, nucleosynthesis, abundances

∗ NSF GRFP Fellow
† Hubble Fellow
‡ Hubble and Carnegie-Dunlap Fellow
2 C HORNOCK ET AL .

1. INTRODUCTION opportunity to test this scenario. An optical and infrared

Studies of the solar system abundance patterns more than counterpart associated with the host galaxy NGC 4993 was
60 years ago demonstrated the existence of a large number independently detected by several teams within the next 12
of heavy isotopes that could not be produced during normal hours (LIGO Scientific Collaboration and Virgo Collabora-
stellar fusion processes and instead required many neutron tion 2017b), including ours (Allam et al. 2017). This optical
captures onto heavy nuclei on timescales short relative to transient source was variously given the name of SSS17a
the weak decay timescale of a few seconds (Burbidge et al. (Coulter et al. 2017b,a) and DLT17ck (Yang et al. 2017;
1957; Cameron 1957). The sites of this phenomenon, given Valenti et al. 2017), as well as an International Astronomical
the name r-process nucleosynthesis, have remained obscure. Union name of AT2017gfo.
Initial efforts focused on supernovae being the best possible Immediately following the detection of a transient optical
candidates then known for the production of the large neutron source by our DECam program (Allam et al. 2017; Soares-
fluxes necessary for the r-process to occur, but most mod- Santos et al. 2017), we triggered our dedicated follow-up
ern work has found that the conditions in numerical simula- spectroscopy program using Gemini-South to search for the
tions are inadequate to produce the heaviest r-process mate- direct signatures of r-process opacity, the detection of which
rial in the necessary quantities (e.g., Qian & Woosley 1996; are reported here. In Section 2, we present the observa-
Thielemann et al. 2011). Compact object mergers involving tions. We perform comparisons to a suite of kilonova mod-
neutron stars were subsequently proposed to be alternative els in Section 3, and conclude in Section 4. Throughout
sites that could easily explain the presence of neutron-rich this work, we assume a distance of 39.5 Mpc and a redshift
material (Lattimer & Schramm 1974; Eichler et al. 1989; of z=0.009727 for NGC 4993 (de Vaucouleurs et al. 1991;
Freiburghaus et al. 1999). Freedman et al. 2001), as well as a correction of E(B − V ) =
Li & Paczyński (1998) noted that synthesis of r-process 0.105 mag for Galactic extinction (Schlafly & Finkbeiner
material would naturally produce a large number of ra- 2011) using the reddening law of Fitzpatrick (1999).
dioactive isotopes whose decay energies could power an
optical transient associated with a neutron star merger (see 2. OBSERVATIONS
also Rosswog 2005). This event has become known as a We used FLAMINGOS-2 on the 8 m Gemini-South tele-
"macronova" (Kulkarni 2005) or "kilonova" (Metzger et al. scope (F2; Eikenberry et al. 2012) to obtain a sequence of
2010). 7 NIR spectra from 1.5 to 10.5 d after the GW trigger. All
Early work assumed that the r-process material would have observations used the same setup with the JH grism and JH
opacities similar to that of iron-peak elements due to a lack filter to cover the wavelength range of 0.9–1.8 µm, at an aver-
of appropriate atomic data (e.g., Metzger et al. 2010). This age dispersion of ∼6.5 Å per pixel. A 4-pixel (0.7200 ) slit re-
assumption works reasonably well for the light r-process nu- sulted in a mean resolution of R ≈ 500 (600 km s−1 ), although
clei (atomic mass number A . 140). However, the lanthanide that figure varies significantly across the spectra. Because of
and actinide series of elements, which are produced in the the high airmass of the observations, we were careful to ori-
r-process, fundamentally differ from other elements in hav- ent the long slit at a position angle of 290◦ , which was very
ing their outer valence electrons in the f -shell. This elec- close to the parallactic value (Filippenko 1982). This was
tronic configuration permits a much larger number of tran- also reasonably close to perpendicular to the offset from the
sitions within a few eV of the ground state, corresponding host galaxy, which minimized the amount of galaxy light in
to a greatly enhanced opacity to optical photons in material the background. The individual exposures were dithered in
enriched in lanthanides (Kasen et al. 2013). Therefore, r- an ABBA sequence, with typical separations of 3000 between
process-enhanced material produced in a kilonova will have the A and B positions. As the source was setting at the begin-
its flux pushed into the near-infrared (NIR) relative to the op- ning of each night, the observations usually started in twilight
tical peak of material whose opacity is dominated by iron- and continued until very high airmass. A log of observations
peak elements (Kasen et al. 2013; Barnes & Kasen 2013; is given in Table 1.
Tanaka & Hotokezaka 2013). While the treatment of the The two-dimensional frames were processed using stan-
atomic data remains highly uncertain, there is general agree- dard procedures in the gemini IRAF1 package to perform
ment that this spectral distinction is a unique signature of dark subtraction and apply a flat field correction. A local sky
material with enhanced r-process abundances (Kasen et al. exposure was created from neighboring dithered frames and
2017; Wollaeger et al. 2017; Tanaka et al. 2017).
The detection of the binary neutron star merger GW170817
1 IRAF is distributed by the National Optical Astronomy Observatory,
by Advanced LIGO/Virgo at 12:41 UT (all times in this paper
which is operated by the Association of Universities for Research in Astron-
are UT) on 2017 August 17 (LIGO Scientific Collaboration omy, Inc., under cooperative agreement with the National Science Founda-
and Virgo Collaboration 2017c,a) afforded an extraordinary tion.
 NIR S PECTROSCOPY OF EM C OUNTERPART TO GW170817 3

Table 1. Log of F2 Spectroscopic Observations

UT Date Epoch Exposure Time Airmass

(mean) (days) (s) Range Age in days after GW170817
⊕
2017 Aug 19.03 1.5 14 × 180 1.8 − 2.6 +1.5
2017 Aug 19.99 2.5 8 × 180 1.4 − 1.7 10
2017 Aug 21.99 4.5 8 × 120 1.5 − 1.7 +2.5

Flux (fλ, 10−17 erg cm−2 s−1 Å−1)

2017 Aug 25.00 7.5 12 × 180 1.6 − 2.1
+4.5
2017 Aug 26.01 8.5 20 × 180 1.7 − 3.0
2017 Aug 27.01 9.5 20 × 180 1.7 − 2.8
2017 Aug 28.01 10.5 16 × 180 1.9 − 2.8
+7.5
(x1.3)

+8.5
then scaled and subtracted prior to registration and combina- 1
tion of the images to form the final spectral stack. A clear +9.5
trace from the transient is present at all epochs, even as the
contrast with the host galaxy star light became small at later
epochs. The host galaxy contributed a significant amount +10.5
of emission along the slit, so a local linear background was
subtracted during spectral extraction. The extracted and
wavelength-calibrated spectra were flux calibrated and cor-
rected for telluric absorption using observations of A0V tel- 0.8 1.0 1.2 1.4 1.6 1.8
luric standard stars observed at a similar airmass to the ob- Rest Wavelength (microns)
ject and the xtellcor_general task in the Spextool IDL
package (Vacca et al. 2003; Cushing et al. 2004).
A pair of H images were taken as part of the acquisition Figure 1. NIR spectral sequence of GW170817 from Gemini-
process for each spectral epoch. The photometry from those South. Each epoch is labeled with its age in days after the GW
trigger. The spectra have been de-redshifted, corrected for Galactic
images is presented by Cowperthwaite et al. (2017). We in-
extinction, and scaled to match the H photometry (Cowperthwaite
tegrated the spectra over the H bandpass and scaled them to et al. 2017). The +7.5 d spectrum has been scaled by an additional
match the photometry. Our final NIR spectral sequence is factor of 1.3 for presentation purposes. The first three spectra are
presented in Figure 1. presented unbinned, but the later ones are binned by increasingly
The initial epoch of NIR spectroscopy, at 1.5 d after the larger factors. The region of strong telluric absorption between J
merger, is very smooth. There is a flattening near 1 µm and a and H bands is indicated by the gray box.
broad shelf present near 1.3–1.4 µm, which unfortunately co-
incides with the strong H2 O vapor absorption band between 3.1. Finding a fiducial model
J and H bands. Only one night later, a peak near 1.05 µm has To proceed further, we need to consult detailed spectral
become prominent and moves redward over the course of the synthesis models. We compare to the library of kilonova
next several nights. By 4.5 d after the merger, a second peak spectral models produced by Kasen et al. (2017), which are
near 1.55 µm develops. This change in appearance, from a an evolved version of previous work (Kasen et al. 2013;
smooth blue continuum to a redder continuum with several Barnes & Kasen 2013). These models assume spherical
bumps in addition to the two dominant peaks, is suggestive symmetry, local thermodynamic equilibrium, and uniform
of a transition in opacity sources of the dominant spectral abundances, with a density structure that has a central core
component over that time. As the transient fades further, the (density ∝ v−1 ) and envelope with a steep density gradient
spectral peaks become less distinct, although that could be (∝ v−10 ). The only three tunable parameters
a consequence of the decreased signal-to-noise ratio and in- p are an ejecta
mass (m), a mean velocity (defined as v = 2K/m, where K
creased host galaxy contamination over time. is the kinetic energy), and a fractional lanthanide abundance
(Xlan ). These models were computed for 56 different com-
3. DISCUSSION binations of these three parameters, with the range of vari-
4 C HORNOCK ET AL .

ation motivated by ejecta parameters of various neutron star and blueward of 1.0 µm that are not visible in the data. We
merger simulations from the literature, such as having ejecta call this the fiducial model because the excellent agreement
masses around a few hundredths of a solar mass. We refer between theory and data implies that whatever more com-
the reader to Kasen et al. (2017) for more details. plicated scenarios we consider below or can be theoretically
We take this library of spectral models, extract the output proposed, the net effect is to produce emission similar to this
at the epochs of our NIR spectra, and convert the model lumi- single-component model.
nosities to fluxes at our adopted distance. All model spectra The basic features of peaks near 1.1 and 1.55 µm, with a
presented in this paper have been smoothed by a 3-pixel box- shelf near 1.25 µm, are also visible in the favored toy model
car to reduce Monte Carlo sampling noise. All intrinsic spec- in the original work of Kasen et al. (2013), although those
tral features in the data are much broader than this smoothing authors also demonstrated that their result depended on the
kernel. underlying atomic structure model. The contrast between
We make a distinction in subsequent discussion between these features and the expected spectral peak at optical wave-
possible "blue" kilonova components, which are relatively lengths for material dominated by iron-peak opacities was re-
lanthanide poor (Xlan .10−4 ) and "red" kilonova components garded as a "smoking gun" of r-process nucleosynthesis. The
(Xlan & 10−3 ). In comparison to neutron star merger simula- models of Tanaka & Hotokezaka (2013) have some roughly
tions, the blue component can represent lanthanide-poor ma- similar features, although their spectra are much broader and
terial ejected primarily in the polar directions at the interface smoother than in our data and show a much larger drop in
between the colliding neutron stars or due to neutrino irradia- flux from 1.0 to 1.6 µm than the data, which may reflect lim-
tion from the newly-formed hypermassive neutron star (e.g., itations of the NIR line list used in that work.
Wanajo et al. 2014; Metzger & Fernández 2014; Goriely et al. We note that independent sets of kilonova models (Wol-
2015). The red component is likely to represent material in laeger et al. 2017; Tanaka et al. 2017), based on new atomic
the equatorial plane, which could originate in tidal tails of structure calculations, also reproduce the shift of the ob-
the disrupted neutron star, spiral mode instabilities, or winds served flux to the NIR when there are high concentrations of
from an accretion disk (e.g., Rosswog et al. 1999; Bauswein lanthanides. The agreement between various codes about this
et al. 2013; Siegel & Metzger 2017). general trend gives us confidence that this signature of the
We start by examining the 4.5 d spectrum. We choose this opacities of r-process elements is robust. However, inspec-
epoch because it has a relatively high signal-to-noise ratio, tion of the figures in those works reveals no clear matches to
has clear spectral features present, is sufficiently late that the the spectral sequence as close as the ones we present here.
early emission from the blue kilonova component has faded The detailed results of those calculations depend on the as-
(Cowperthwaite et al. 2017; Nicholl et al. 2017), and yet is sumptions about the masses and compositions of different
well before the ejecta become optically thin and the radia- merger ejecta components. It is unclear at this point whether
tive transfer assumptions might break down. A simple least- these detailed spectral differences from different codes rep-
squares statistic finds a best fit among available models with resent alternative assumptions about the parameters of the
one that has a total mass of 0.04 M , expansion velocity 0.1c, neutron star merger ejecta or differences in the treatment of
and a fractional lanthanide abundance Xlan of 10−2 , which is opacities across the lanthanide series.
similar to the properties of the red kilonova component in-
dependently inferred from the optical and NIR light curves 3.2. Sensitivity to Parameters
by Cowperthwaite et al. (2017). A comparison between the
We take the excellent agreement between model and data
model and data is shown in Figure 2.
shown in Figure 2 as a sign that the parameters and the mod-
The agreement between model and data is astonishingly
els are at least roughly correct, so now we examine the sen-
good, especially given the uncertainties in the underly-
sitivity of the model output to the parameter values we have
ing atomic data, the simplifications inherent to a single-
selected. In the three panels of Figure 3, we vary each of the
component model, and the lack of any previous NIR spectra
three main parameters in sequence, while holding the other
for guidance. We emphasize that we have applied no arbi-
two fixed. Each of the model spectra also includes a small
trary flux scaling factors to the model, and yet it matches the
amount of flux from the same assumed blue kilonova com-
height and wavelength of the 1.07 µm peak very well. There
ponent (not shown, but parameters are discussed below) that
is a peak in the model spectrum near 1.5 µm that is some-
contributes a small amount of flux below 1 µm.
what bluer and higher than the one in the data. Some of the
In the top panel, we start by varying the ejecta mass. As
less prominent features, such as the two shelves of emission
we lower the mass, the overall flux goes down, as expected.
between 1.1 and 1.25 µm on the red side of the 1.07 µm
However, the spectra are not simply related by a flux nor-
peak, are also present in both spectra. The largest disagree-
malization factor. The total mass in the ejecta also affects
ment is in the presence of a pair of dips in the model near
the diffusion timescale and hence the location of the photo-
 NIR S PECTROSCOPY OF EM C OUNTERPART TO GW170817 5

7
GW170817 Kasen et al. 2017
+4.5 d red kilonova model
6
fλ (10−17 erg cm−2 s−1 Å−1)

M = 0.04 MO•
⊕
v = 0.1c
5 Xlanthanide = 10-2

1.0 1.2 1.4 1.6 1.8

Rest Wavelength (microns)
Figure 2. The fiducial red kilonova model provides an excellent fit by itself to the day +4.5 NIR spectrum, with no adjustments to the flux
scale. The data are in black and the model is in red, with the values of the three main parameters listed on the figure.

sphere within the ejecta. This, in turn, results in variations (e.g., Kasen et al. 2015; Siegel & Metzger 2017) predict a
in the amount of line blending that shift the wavelengths of range of ejection velocities from 0.03–0.1c. We do not see
the spectral peaks. Most notably, the 1.07 µm peak shifts narrow features expected from material moving as slowly as
redward at lower ejecta mass. v = 0.03c at any epoch. If the red kilonova ejecta result from
The ejecta velocity also affects the degree of line blend- a disk wind, they must be accelerated above this value by,
ing and smoothness of the spectra. In the middle panel of for example, stronger magnetic fields than those previously
Figure 3, it is clear that raising the ejecta velocity rounds the considered.
tops of the spectral peaks and at v = 0.2c, the features between Finally, the most important question for the purposes of r-
1.1–1.3 µm are unacceptably washed out relative to the data. process nucleosynthesis is constraining the chemical abun-
We note that some simulations of the tidal dynamical ejecta dances of the dominant emission component. In the bot-
find even higher ejecta velocities than this (e.g., Bauswein tom panel of Figure 3, we have adjusted the fractional lan-
et al. 2013). At the other extreme, lowering the ejection ve- thanide abundance. If the lanthanide abundance is as low as
locity results in the major peaks breaking up into a forest of Xlan =10−4 , the peak near 1.1 µm disappears and the model
smaller peaks. The v = 0.03c spectrum presented in this panel spectra are too blue. At abundances that are much higher
shows several of these features starting to develop. Although than our fiducial model, the peak near 1.1 µm is suppressed
it is not plotted, by 7.5 d these narrower peaks are predicted relative to the one near 1.5 µm. Models with Xlan between
to get even more dominant, in contradiction to the smooth 10−2 and 10−3 appear to match the overall appearance and ra-
broad peaks that we see at that time (Figure 1). This is rel- tio of peak heights reasonably well. The overall flux scaling
evant because models invoking strong accretion disk winds
6 C HORNOCK ET AL .

al. 2017) that at the earliest times emission at optical wave-

fλ (10−17 erg cm−2 s−1 Å−1) M/MO• v/c Xlan
8 +4.5 d lengths is dominated by a lanthanide-poor and high-velocity
0.040 0.10 1e-2
0.035 0.10 1e-2 blue kilonova component.
6 ⊕ 0.030 0.10 1e-2
0.025 0.10 1e-2 This observational result, plus theoretical guidance (e.g.,
4 Metzger & Fernández 2014), motivates us to consider two-
component kilonova models that have both blue and red com-
2 ponents with different r-process abundances. We shall pro-
Varying mass ceed by simply summing the flux produced by two separate
0
1.0 1.2 1.4 1.6 1.8 single-component models. This is clearly a questionable as-
Rest Wavelength (microns) sumption, but can be justified if either the two kilonova com-
ponents are spatially disjunct (e.g., a blue polar flow and a
fλ (10−17 erg cm−2 s−1 Å−1)

8 M/MO• v/c Xlan red equatorial component) or if one component has a much
0.040 0.10 1e-2
0.025 0.03 1e-2 higher velocity than the other. We will revisit this below.
6 ⊕ 0.040 0.15 1e-2 The best simple summation of single-component models
0.040 0.20 1e-2
relative to the first four spectral epochs that we have found is
4
shown in Figure 4. This model has both a blue (Xlan =10−5 )
2 and a red (Xlan =10−2 ) component, whose sum is shown in
Varying velocity purple on each panel of the figure. While it is not as good a
0 fit to the 4.5 d spectrum as the fiducial model, it alleviates the
1.0 1.2 1.4 1.6 1.8
Rest Wavelength (microns) discrepancies with the flux levels at earlier and later epochs
described above.
We start by examining the 1.5 d spectrum because the
fλ (10−17 erg cm−2 s−1 Å−1)

8 M/MO• v/c Xlan

0.040 0.10 1e-2.0 blue kilonova component only makes a dominant contribu-
1e-1.5
6 1e-2.5 tion at early times. The abundances and velocity (0.3c) of
⊕
1e-3.0 the blue component are the same found from fitting the opti-
1e-4.0
4 cal photometry and spectroscopy (Cowperthwaite et al. 2017;
Nicholl et al. 2017), and with an ejecta mass (0.02 M ) be-
2
tween those two estimates. However, this model still has
Varying lanthanide fraction some inconsistencies. The actual data are much smoother
0
1.0 1.2 1.4 1.6 1.8 than the combined model, implying that the photosphere is
Rest Wavelength (microns)
in high-velocity material with blue colors. A pure blue kilo-
nova model with a higher ejecta mass (as found by Nicholl et
Figure 3. The effect of varying model parameters. Top panel: vary-
al. 2017) would probably be a better fit by itself. However,
ing the ejecta mass. Middle panel: the ejecta velocity. The low-
velocity v = 0.03c model also has a lower ejecta mass to keep the any red kilonova component sufficiently luminous to account
flux level consistent with the others. Bottom panel: the mass frac- for the later spectra also contributes some emission at these
tion of lanthanide elements. In each panel, the data are shown in early times as well. In addition, this red component imprints
black and our fiducial model is red. The text labels in each panel a spectral dip near 1.25 µm on the combined model that is not
give the model parameters for the lines of the same color. seen in the data. One possible resolution of these inconsis-
tencies is if while the photosphere is in the high-velocity blue
at lower lanthanide fraction can be somewhat improved by material at early times, it obscures the contribution from the
lowering the ejecta mass. lower-velocity red kilonova component until the blue kilo-
nova becomes more optically thin at later epochs.
3.3. Spectral evolution Next, we compare to the later-time spectra (7.5 d). As
described above, the fiducial red kilonova model provides
The agreement between data and the model in Figure 2,
an excellent fit to the 4.5 d spectrum, but over-predicts the
while impressive, is only part of the story. The next challenge
late-time flux. The minimal modifications to the fiducial
for the models is to self-consistently match the time evolution
model to resolve this are to reduce the radiative diffusion
in the full spectral sequence.
timescale by some combination of reducing the ejecta mass
We find that the fiducial model parameters from Figure 2
(here, 0.035 M ) and increasing the velocity (0.15c). The
overstate the NIR flux at 7.5 d by almost a factor of 2 and
flux level at 7.5 d is now acceptable. However, all model cal-
under-predict the flux at 1.5 d. The latter is not a surprise, as
culations we examined with Xlan =10−2 develop a broad third
there are independent lines of evidence from the optical pho-
NIR peak between 1.2–1.3 µm at 7.5 d that is not present in
tometry (Cowperthwaite et al. 2017) and spectra (Nicholl et
 NIR S PECTROSCOPY OF EM C OUNTERPART TO GW170817 7

Rest Wavelength (microns) Rest Wavelength (microns)

1.0 1.2 1.4 1.6 1.8 1.0 1.2 1.4 1.6 1.8
12
GW170817 +1.5 d +2.5 d
fλ (10−17 erg cm−2 s−1 Å−1)

fλ (10−17 erg cm−2 s−1 Å−1)

15 10
⊕ blue kilonova component red kilonova component
M = 0.020 MO• M = 0.035 MO• 8
⊕ v = 0.15c
v = 0.30c
10 Xlanthanide = 10-2.0
Xlanthanide = 10-5.0 6

4
5
2
0 0
+4.5 d +7.5 d
fλ (10−17 erg cm−2 s−1 Å−1)

fλ (10−17 erg cm−2 s−1 Å−1)

3
6 ⊕
⊕
2
4

2 1

0 0
1.0 1.2 1.4 1.6 1.8 1.0 1.2 1.4 1.6 1.8
Rest Wavelength (microns) Rest Wavelength (microns)

Figure 4. A combined blue and red kilonova model that is our best simultaneous fit to the first four epochs of F2 spectra. The data are shown
in black. A blue kilonova model is shown in blue and the red kilonova component is in red, with the parameters for each model listed in the
upper two panels. The thick purple line represents the sum of the two components.

the data. Reducing the lanthanide concentration to Xlan =10−3 dominant opacity sources. Our highest-quality spectrum is a
produces spectra with only two peaks in the NIR at this epoch good match to a fiducial one-component red kilonova model
(at 1.1 and 1.6 µm), in agreement with the data. This may from Kasen et al. (2017), despite many theoretical uncertain-
be similar to the intermediate or "purple" kilonova compo- ties in the atomic data and heating rate (e.g., Barnes et al.
nent inferred from the photometry by Cowperthwaite et al. 2016; Hotokezaka et al. 2016). We find that in order to re-
(2017). However, such a model is too blue at 2.5 d and lacks produce the spectral peaks near 1.05 µm and 1.55 µm that
a spectral peak near 1.05 µm as strong as that seen in the dominate the spectra as soon as 2.5 d after the merger, we
data. require kilonova ejecta with a mass of ∼0.04 M and a high
In summary, we have experimented with resolving the fail- lanthanide abundance, Xlan ≈10−2 . Models that lack a suffi-
ures of a single-component kilonova model to match the cient concentration of material with lanthanide-like opacities
spectral evolution by summing a blue and a red component. result in spectra that are too blue and otherwise inconsistent
We present such a model in Figure 4 and find that the overall with the data.
flux level is reasonably good, but that there are several in- In order to better match the spectral sequence, we also ex-
consistencies. We also note that if we were to fit the first four plored simply adding blue and red kilonova components with
spectra independently, we would find best-fit lanthanide frac- different lanthanide abundances. The properties of the blue
tions of . 10−4.5 , 10−2 , 10−2 , and 10−3 for days 1.5, 2.5, 4.5, component are similar to those found to fit the optical spec-
and 7.5, respectively. No combination of two (or three) com- tra (Nicholl et al. 2017). In addition, the presence of at least
ponents can reproduce this evolution by simple summation, two components with parameters similar to ours was also in-
so more complex models are required. The effects of geome- ferred from fitting the combined optical and NIR light curves
try, self-obscuration, and reprocessing of emission from one (Cowperthwaite et al. 2017).
component by the other are likely important in reality, but Further progress will require models that have more real-
have not been treated in these comparisons. istic variations of abundances with velocity within the ejecta.
The numbers quoted at the end of Section 3.3 for the lan-
4. CONCLUSION thanide abundances in the dominant emission component at
We have presented a spectral sequence of the NIR coun- each epoch may provide a guide for future work. The excel-
terpart to the binary neutron star merger GW170817. Our lent agreement between model and data shown in Figure 2
initial spectra are smooth and blue, but develop several broad may be a sign that the properties of the component that dom-
peaks starting 2.5 d after the merger, indicating a change in
8 C HORNOCK ET AL .

inates the emission at 4.5 d has parameters similar to those of gentina), and Ministério da Ciência, Tecnologia e Inovação
the model. (Brazil).
Matching the derived Galactic r-process production rate The Berger Time-Domain Group at Harvard is supported
requires kilonova ejecta masses in the range of 0.01–0.1 M in part by the NSF through grants AST-1411763 and AST-
(e.g., Fernández & Metzger 2016) if estimates of the binary 1714498, and by NASA through grants NNX15AE50G and
neutron star merger rate from the Galactic neutron star pop- NNX16AC22G.
ulation are correct (Kalogera et al. 2004; Kim et al. 2015). D.K. is supported in part by a Department of Energy Of-
Our best fit two-component model has a total ejecta mass of fice of Nuclear Physics Early Career Award and by grant DE-
0.055 M , with 0.035 M of material enriched in the heavy SC0017616, and by the Director, Office of Energy Research,
r-process, squarely in the middle of the range. Further im- Office of High Energy and Nuclear Physics, Divisions of Nu-
provements in the merger rate from Advanced LIGO/Virgo clear Physics, of the U.S. Department of Energy under Con-
detections and observational constraints on the range of kilo- tract No.DE-AC02-05CH11231. D.K.’s research was sup-
nova ejecta masses will greatly reduce the uncertainty in ported in part by the Gordon and Betty Moore Foundation
these estimates. Future observations using the James Webb through Grant GBMF5076. This research used resources of
Space Telescope will take advantage of the broad spectral the National Energy Research Scientific Computing Center, a
range and lack of telluric absorption to enable more com- DOE Office of Science User Facility supported by the Office
plete studies of the merger ejecta components and their de- of Science of the U.S. Department of Energy under Contract
tailed abundance patterns. However, as long as GW170817 No. DE AC02-05CH11231.
is typical of the population, our results demonstrate that bi- D.A.B. is supported by NSF award PHY-1707954.
nary neutron star mergers do produce sufficient quantities of R.J.F. thanks the University of Copenhagen, DARK Cos-
r-process material. mology Centre, and the Niels Bohr International Academy
for hosting him during the discovery of GW170817/SSS17a,
where he was participating in the Kavli Summer Program in
Astrophysics, "Astrophysics with gravitational wave detec-
tions." This program was supported by the Kavli Founda-
We are grateful for the heroic efforts of the entire Gemini- tion, Danish National Research Foundation, the Niels Bohr
South staff to obtain these observations, which started in twi- International Academy, and the DARK Cosmology Cen-
light each night as the source was setting, particularly M. An- tre. The UCSC group is supported in part by NSF grant
dersen, P. Candia, J. Chavez, G. Diaz, R. Diaz, V. Firpo, G. AST–1518052, the Gordon & Betty Moore Foundation, the
Gimeno, H. Kim, A. Lopez, L. Magill, P. Prado, R. Rutten, R. Heising-Simons Foundation, generous donations from many
Salinas, D. Sanmartim, A. Shugart, K. Silva, E. Wenderoth, individuals through a UCSC Giving Day grant, and from fel-
and the director, L. Ferrarese, for her support and approval of lowships from the Alfred P. Sloan Foundation and the David
the associated DD program. and Lucile Packard Foundation to R.J.F.
Based on observations obtained at the Gemini Observa- M.R.D. is supported by NASA through Hubble Fellow-
tory (Program IDs GS-2017B-Q-8 and GS-2017B-DD-4; PI: ships awarded by the Space Telescope Science Institute,
Chornock), which is operated by the Association of Univer- which is operated by the Association of Universities for Re-
sities for Research in Astronomy, Inc., under a cooperative search in Astronomy, Inc., for NASA, under contract NAS
agreement with the NSF on behalf of the Gemini partnership: 5-26555.
the National Science Foundation (United States), the Na-
tional Research Council (Canada), CONICYT (Chile), Min-
Facilities: Gemini:South (FLAMINGOS-2)
isterio de Ciencia, Tecnología e Innovación Productiva (Ar- Software: IRAF, IDL, Spextool

REFERENCES
Allam et al. 2017, GRB Coordinates Network, 21530 Coulter et al. 2017a, Science, Accepted
Barnes, J., & Kasen, D. 2013, ApJ, 775, 18 —. 2017b, GRB Coordinates Network, 21529
Barnes, J., Kasen, D., Wu, M.-R., & Martínez-Pinedo, G. 2016,
Cowperthwaite et al. 2017, ApJL, Submitted
ApJ, 829, 110
Cushing, M. C., Vacca, W. D., & Rayner, J. T. 2004, PASP, 116,
Bauswein, A., Goriely, S., & Janka, H.-T. 2013, ApJ, 773, 78
362
Burbidge, E. M., Burbidge, G. R., Fowler, W. A., & Hoyle, F.
1957, Reviews of Modern Physics, 29, 547 de Vaucouleurs, G., de Vaucouleurs, A., Corwin, Jr., H. G., et al.
Cameron, A. G. W. 1957, PASP, 69, 201 1991, Third Reference Catalogue of Bright Galaxies.
 NIR S PECTROSCOPY OF EM C OUNTERPART TO GW170817 9

Eichler, D., Livio, M., Piran, T., & Schramm, D. N. 1989, Nature, —. 2017b, ApJL, in prep.
340, 126 —. 2017c, GRB Coordinates Network, 21509
Eikenberry, S., Bandyopadhyay, R., Bennett, J. G., et al. 2012, in Metzger, B. D., & Fernández, R. 2014, MNRAS, 441, 3444
Proc. SPIE, Vol. 8446, Ground-based and Airborne
Metzger, B. D., Martínez-Pinedo, G., Darbha, S., et al. 2010,
Instrumentation for Astronomy IV, 84460I
MNRAS, 406, 2650
Fernández, R., & Metzger, B. D. 2016, Annual Review of Nuclear
and Particle Science, 66, 23 Nicholl et al. 2017, ApJL, Submitted
Filippenko, A. V. 1982, PASP, 94, 715 Qian, Y.-Z., & Woosley, S. E. 1996, ApJ, 471, 331
Fitzpatrick, E. L. 1999, PASP, 111, 63 Rosswog, S. 2005, ApJ, 634, 1202
Freedman, W. L., Madore, B. F., Gibson, B. K., et al. 2001, ApJ, Rosswog, S., Liebendörfer, M., Thielemann, F.-K., et al. 1999,
553, 47 A&A, 341, 499
Freiburghaus, C., Rosswog, S., & Thielemann, F.-K. 1999, ApJL, Schlafly, E. F., & Finkbeiner, D. P. 2011, ApJ, 737, 103
525, L121
Siegel, D. M., & Metzger, B. D. 2017, ArXiv e-prints,
Goriely, S., Bauswein, A., Just, O., Pllumbi, E., & Janka, H.-T.
arXiv:1705.05473 [astro-ph.HE]
2015, MNRAS, 452, 3894
Hotokezaka, K., Wanajo, S., Tanaka, M., et al. 2016, MNRAS, Soares-Santos et al. 2017, ApJL, Submitted
459, 35 Tanaka, M., & Hotokezaka, K. 2013, ApJ, 775, 113
Kalogera, V., Kim, C., Lorimer, D. R., et al. 2004, ApJL, 601, L179 Tanaka, M., Kato, D., Gaigalas, G., et al. 2017, ArXiv e-prints,
Kasen, D., Badnell, N. R., & Barnes, J. 2013, ApJ, 774, 25 arXiv:1708.09101 [astro-ph.HE]
Kasen, D., Fernández, R., & Metzger, B. D. 2015, MNRAS, 450, Thielemann, F.-K., Arcones, A., Käppeli, R., et al. 2011, Progress
1777 in Particle and Nuclear Physics, 66, 346
Kasen, D., Metzger, B., Barnes, J., Quataert, E., & Ramirez-Ruiz,
Vacca, W. D., Cushing, M. C., & Rayner, J. T. 2003, PASP, 115,
E. 2017, Nature, Submitted
389
Kim, C., Perera, B. B. P., & McLaughlin, M. A. 2015, MNRAS,
448, 928 Valenti et al. 2017, ApJ, accepted
Kulkarni, S. R. 2005, ArXiv Astrophysics e-prints, Wanajo, S., Sekiguchi, Y., Nishimura, N., et al. 2014, ApJL, 789,
astro-ph/0510256 L39
Lattimer, J. M., & Schramm, D. N. 1974, ApJL, 192, L145 Wollaeger, R. T., Korobkin, O., Fontes, C. J., et al. 2017, ArXiv
Li, L.-X., & Paczyński, B. 1998, ApJL, 507, L59 e-prints, arXiv:1705.07084 [astro-ph.HE]
LIGO Scientific Collaboration and Virgo Collaboration. 2017a, Yang et al. 2017, GRB Coordinates Network, 21531
Phys. Rev. Lett., in prep.